Polyphony Digital Helps New AI Deep Learning Study Using Gran Turismo

  • Thread starter Famine
  • 26 comments
  • 3,268 views

zzz_pt

Premium
8,125
Germany
Porto living in Hamburg
zzz_pt
Great news if sometime in the near future we can see it implemented in GT. 3,3sec slower than the top10% is still slow for me but progress is being made, hopefully.

It’s very good the AI is based on human like behaviour and limited to the same physics instead of artificially increasing the pace beyond what the car/tires/surface allows as some games (including GT) often implement.
 
Last edited:
72
Guatemala
Guatemala
It’s very good the AI is based on human like behavior and limited to the same physics instead of artificially increasing the pace beyond what the car/tires/surface allows as some games (including GT) often implement.
This for me would be its most impressive and desired feature added to a GT game. I really hate rubberbanding as it makes the game arcadish and we already have tons of games like that. If Gran Turismo is to be called "a real driving simulator" then enhance its artificial capabilities. This also applies to the extremely slow AI in GT Sport. Even in Professional mode it's too slow and boring.
 
78
Spain
Gijon
Coquico
Great news if sometime in the near future we can see it implemented in GT. 3,3sec slower than the top10% is still slow for me but progress is being made, hopefully.

It’s very good the AI is based on human like behaviour and limited to the same physics instead of artificially increasing the pace beyond what the car/tires/surface allows as some games (including GT) often implement.
Exactly. For me, Beside the lap time, what the game need its a more human AI behavior. Not breaking in the midle of the turn and then existing it like an harpoon missile, for example. The gt AI has bad behavior long time ago. This improvement its in my wishlist!
 
39,143
Australia
The Bronx
That’s pretty good for a basic car on a tight circuit. Try the LaFerrari at DTG next, please. ;)
 
1,565
Canada
Montreal, Canada
NobleAtreides
B- Spec????

🥰🥰🥰🥰🥰😍
Ai - Spec !
OK PS5, when can it be my turn to play?..

PS5: oh human, if only you can learn at the pace that I suddenly am ! (self centers wheel, & spins frowzily while drifting the entire nurburgring in reverse)
 
Last edited:
56
United Kingdom
United Kingdom
w_a_i_n
Surely "perfect AI", is player dependent. If you have "alien drivers", they need harder AI that can match and push them.

If casual gamers enjoy GT, but these players can't be fast with pad/wheel, then they need slower AI. In reality, casual GT drivers may spend less time hotlapping and perfecting lines/setups etc, and use other features like liveries, scapes, etc. Also casual players may often prefer some assists, which whilst helpful, it makes them slower too.

Whilst any AI improvement work is great, this stuff has to not penalty those who own lots of games, but likely don't drive as much as those in the top 50%. Modern consoles are easily capable of taking the times set by players in their own driving sessions, and tuning the experience to match them. Even to the point, that if one player likes American tracks, and is faster on those tracks, the AI can adjust for that too.

This would also IMHO, would be a fairer system for the 3 "medals" earned. I know people who can't drive and win Gold's in the challenges - which is gutting for them tbh. If GT "tuned" the AI, for the players own skill/experience, it could set medal benchmark times based on the individual. Settings could also then be changed from DEFAULT=PRO/MEDIUM/EASY/ETC to AI.LEARNING=BEATME/MATCHME/LETMEWINGODDAMMIT.

This might sound complex, but it needn't be. For example, you could have an AI lap which can drive and react much much faster than the player easily - BUT, use a variable to "adjust" that lap to bring the total laptime down to within range. There can be lots of stored data on the players system, which can be used by the AI code to help tune the pace of the AI. For a day-one experience, players could simply resort to the set menu options for difficulty until enough reference laps have been stored in the game. The laps could be supplied by licence tests, SP races, etc.

Interesting times ahead!!

:)
 
1,401
This AI will allow PDI to replace Bob Weaver as their default AI driver. It would allow them to dynamically adjust its competitiveness level during career events.

The next step is to get the AI strategy right, rather than just relying on a script so that we can't lap the AI or finish 2 minutes ahead of second place.
 
Last edited:
2,773
Austria
Austria
GTP_SnowCrash
I fail to see any inkling that any of this will find its way into GT7.

Will the AI still shift each car at the red line in GT7, no matter what the power curve would suggest? (If the answer to that is yes, you can already put the PP-system in the bin)

Edit: Maybe a word of explanation - PD says that the PP figure will be determined by the game's AI running the car a number of times around (a) track(s). If the AI is shifting at the wrong times, this will produce slower lap times for certain cars and therefore inflated PP values.
 
Last edited:
39,143
Australia
The Bronx
I fail to see any inkling that any of this will find its way into GT7.

Will the AI still shift each car at the red line in GT7, no matter what the power curve would suggest? (If the answer to that is yes, you can already put the PP-system in the bin)

Edit: Maybe a word of explanation - PD says that the PP figure will be determined by the game's AI running the car a number of times around (a) track(s). If the AI is shifting at the wrong times, this will produce slower lap times for certain cars and therefore inflated PP values.
Will the majority of players(which seem likely to be casuals) shift at the redline no matter what the power curve would suggest?
If casual players are shifting at the wrong times, they will produce slower times for certain cars and therefore even out with similar AI. ;) PD are onto something. Let's wait(we've waited 20 years, another 4-5 won't hurt) and see.
 

Jump_Ace

Compare Yourself To The Best Of Your Own Abilities
Staff Emeritus
8,589
United States
Portland, OR
Jump_Ace
JUMPxACE
I want to be excited for this news, however the AI is really only important to me in the beginning as that is when I spend 99% of my time against the AI. Which means it has to be a day one thing to get the most benefit (and if it's well done I'd come back to the races again later) and I see this a highly unlikely. If Assetto Corsa Competizione can put together solid AI, why can't PD who have been doing this since like '97? shrug


Jerome
 
1,401
I want to be excited for this news, however the AI is really only important to me in the beginning as that is when I spend 99% of my time against the AI. Which means it has to be a day one thing to get the most benefit (and if it's well done I'd come back to the races again later) and I see this a highly unlikely. If Assetto Corsa Competizione can put together solid AI, why can't PD who have been doing this since like '97? shrug


Jerome
Sport's AI are acceptable if you control performance of your car (grip, power and weight) enough that it evens the odds of winning. If your BoP overwhelms the AI, then it lowers the thrill value of not knowing if you will win or not. Sport League events definitely can be set up for having fun.

As far as the abilities or challenge from Sport's AI, if you drive against them enough, you can see them change their habits and skill level as you change the performance of your car. It won't change Bob Weaver's habit of going from side to side of the track all that much, but, still, if you repeat a race often enough, you will clearly see what they do. A driver who is willing to learn this way will improve his skills fairly quickly against the AI which can transfer to Custom on-line races.

As far as GT's AI's quality - meh. It is what it is. We all know the rate of change coming from headquarters. It's not as if we can swap out its AI and replace it with an alternative. Make the best of what we get.
 
3,154
France
France
This for me would be its most impressive and desired feature added to a GT game. I really hate rubberbanding as it makes the game arcadish and we already have tons of games like that. If Gran Turismo is to be called "a real driving simulator" then enhance its artificial capabilities. This also applies to the extremely slow AI in GT Sport. Even in Professional mode it's too slow and boring.
In FH5, AI is just plainly cheating. At least, it's fast but too fast. GT7 without a great AI will be a failure for many players, there is no point with a career mode so far with the GTSport AI (and GT6 and GT5...).
 
2
United States
United States
A lot of optimism in this thread as this has been proven to be a blessing or a curse. Turn10 have been working on this since Motorsport 5 and they've explained how it created issues that resulted in what we see today. I'm curious to see how Polyphony would implement it.
 
Last edited:
3,914
Australia
Australia
Interesting, but I dont see this as anything more than PD allowing GT to be used as a test platform. Its not for PD's direct benefit.

Besides - the game doesn't need the AI to be trained on visual cues - the AI already has access to every parameter of the track.
 
Last edited:

LeGeNd-1

Premium
6,864
Australia
Australia
GTP_LeGeNd-1
Yeah I'm not convinced this is anything other than an academic curiosity to be honest.

No home console is going to be able to run the AI learning simulations live, for 16 cars on track.

That means running the simulations in advance, for every car+track combination in the game. And that's before accounting for weather, tyre/fuel and tuning. And that's just for a car running on its own on track, not driving against human players. It's gonna take a huge amount of time and the algorithm would probably fill the whole game disc on its own.

Other games managed to make fast challenging and reactive AI for years now without resorting to machine learning. It can be done, just a matter of whether PD wants to or not.
 
622
United States
United States
Great news if sometime in the near future we can see it implemented in GT. 3,3sec slower than the top10% is still slow for me but progress is being made, hopefully.

It’s very good the AI is based on human like behaviour and limited to the same physics instead of artificially increasing the pace beyond what the car/tires/surface allows as some games (including GT) often implement.
That was just for one experiment. The first experiment the AI was faster than any human driver.
Yeah I'm not convinced this is anything other than an academic curiosity to be honest.

No home console is going to be able to run the AI learning simulations live, for 16 cars on track.

That means running the simulations in advance, for every car+track combination in the game. And that's before accounting for weather, tyre/fuel and tuning. And that's just for a car running on its own on track, not driving against human players. It's gonna take a huge amount of time and the algorithm would probably fill the whole game disc on its own.

Other games managed to make fast challenging and reactive AI for years now without resorting to machine learning. It can be done, just a matter of whether PD wants to or not.
It wouldn't be that strange to run it for every track. That's basically the essence of "machine learning" which is central to every modern AI model.